Cross-Validation

1/**2 * Example 23: Cross-Validation3 *4 * Cross-validation scores a model on several different train/test splits, so5 * one lucky or unlucky split does not decide the result. This example shows the6 * splitters (KFold, StratifiedKFold, LeaveOneOut) and crossValScore, which runs7 * the whole loop for you.8 */9 10import { crossValScore, LinearRegression } from "deepbox/ml";11import { arange, tensor } from "deepbox/ndarray";12import { KFold, LeaveOneOut, StratifiedKFold } from "deepbox/preprocess";13import { rand, setSeed } from "deepbox/random";14 15// Synthetic linear data: y = 2x + 3 + noise in [-0.5, 0.5). The seed makes it repeatable.16setSeed(42);17const X = arange(0, 50).div(5).reshape([50, 1]);18const y = X.reshape([50])19  .mul(2)20  .add(3)21  .add(rand([50]).sub(0.5));22 23console.log(`Dataset: ${X.shape[0]} samples\n`);24 25// 1. K-Fold: split the rows into k folds. Each fold is the test set once.26console.log("1. K-Fold Cross-Validation (k=5):");27console.log("-".repeat(50));28 29const kfold = new KFold({ nSplits: 5, shuffle: true, randomState: 42 });30 31let foldNum = 1;32for (const { trainIndex, testIndex } of kfold.split(X)) {33  // The splitter yields row indices. Use them to pick rows, for example with X.gather(...).34  console.log(`Fold ${foldNum}: Train=${trainIndex.length}, Test=${testIndex.length}`);35  foldNum++;36}37 38console.log(`\nTotal folds: ${kfold.getNSplits()}\n`);39 40// 2. Stratified K-Fold: every fold keeps the class proportions of y.41console.log("2. Stratified K-Fold:");42console.log("-".repeat(50));43 44// 12 samples, 3 classes with 4 samples each45const yClass = tensor([0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2]);46const XClass = arange(0, 12).reshape([12, 1]);47 48const stratified = new StratifiedKFold({49  nSplits: 2,50  shuffle: true,51  randomState: 42,52});53 54const labels = yClass.toArray() as number[];55foldNum = 1;56for (const { trainIndex, testIndex } of stratified.split(XClass, yClass)) {57  // Count how many test rows belong to each class.58  const counts = [0, 1, 2].map((c) => testIndex.filter((i) => labels[i] === c).length);59  console.log(60    `Fold ${foldNum}: Train=${trainIndex.length}, Test=${testIndex.length}, test rows per class=[${counts}]`61  );62  foldNum++;63}64 65console.log("\nEach test fold holds the same number of rows from every class\n");66 67// 3. Leave-One-Out: n folds, each tests a single row.68console.log("3. Leave-One-Out Cross-Validation:");69console.log("-".repeat(50));70 71const XSmall = tensor([[1], [2], [3], [4], [5]]);72const loo = new LeaveOneOut();73const looFolds = Array.from(loo.split(XSmall));74 75console.log(`Total folds: ${looFolds.length}`);76console.log("Each fold trains on n-1 samples and tests on 1");77console.log("Uses the most training data, but fits the model n times\n");78 79// 4. crossValScore: fit and score a model on every fold in one call.80console.log("4. crossValScore:");81console.log("-".repeat(50));82 83// For a regressor the score is R². The folds come from a fixed-seed shuffle.84const scores = crossValScore(new LinearRegression(), X, y, 5);85console.log(`Scores per fold: ${scores.map((s) => s.toFixed(4)).join(", ")}`);86const meanScore = scores.reduce((a, b) => a + b, 0) / scores.length;87console.log(`Mean R²: ${meanScore.toFixed(4)}\n`);88 89console.log("Summary:");90console.log("  K-Fold: a good default, 5 or 10 folds");91console.log("  Stratified K-Fold: keeps class proportions, use it for classification");92console.log("  Leave-One-Out: maximum training data per fold, high variance, slow on large data");